Hotel Search, Scalability, and Apache Ignite Musaul Karim Senior - PowerPoint PPT Presentation

Hotel Search, Scalability, and Apache Ignite Musaul Karim • Senior Consultant • June 2018

A G E N D A Introduction Hotel Search Systems Architecture Successes & Challenges Questions In-Memory Computing Summit, London • 25-26 June 2018

About Me Software Consultant ● In-memory & Distributed Systems Specialist ● MSc Distributed Computing Initial Career CG Consultancy § 2000 - Started as a C++ developer § IT System Migration Projects § 2003 - Took a break to do my MSc § Technology Assessment § 2005 - Back into world of work at Deloitte § Options within the Modern Landscape § Proof of Concepts In-Memory Systems § Leading Follow up Development Work 2007 - Fidessa § Overall Technical Architecture § High Transaction Order & Execution Management System § In-house developed Distributed Cache Systems for Trade Data 2010 - Barclays Travel sector clients § Migrated DBMS based Risk Calculation engine to an In-Memory § JacTravel Cache & Compute system § OAG § Hybrid In-house tech + Solace Systems + Oracle Coherence 2013 - Credit Suisse § Recently started working with one of the largest travel operators § Oracle Coherence based Prime Services Risk System In-Memory Computing Summit, London • 25-26 June 2018

Hotel Search Systems In-Memory Computing Summit, London • 25-26 June 2018

Hotel Search System Overview § Handles Hotel/Room Search requests via a B2B API § Receives updates intraday as streams as well as batches from Booking Systems and other Third Party Supplier Systems § Returns Priced Rooms matching the Search Criteria § Matches Hotels based on locations searched (Can also search for specific hotels) § Matches Rooms based on Stay Date Availability and Occupancy requirements etc. § Excludes rooms based on any distribution rules § Calculates prices for all the room options § Typically more I/O bound than CPU § It requires a large number of queries against Database Tables (or Caches) at each stage § Large number of calculations to be performed. i.e. they need to be done for each room / special offer / room-extras etc. In-Memory Computing Summit, London • 25-26 June 2018

Search Journey Hotel & Room Selection Cost & Price Calculation Finalise Result (Per Room, Dynamic) Select Hotels Calculate Cost - Location - Contracts - Distribution Rules Deduplicate Rooms Apply Special Offers & Supplements Select Rooms - Room criteria - Occupancy Rules Apply Margin Build and Return Response Filter by Availability - Availability Apply Tax - Stay Period + restrictions In-Memory Computing Summit, London • 25-26 June 2018

Architecture In-Memory Computing Summit, London • 25-26 June 2018

Previous Infrastructure at JacTravel § Two Platforms § One retained as a booking platform (iVector) § The other being decommissioned (TravelSudio). § Built on Microsoft SqlServer and IIS (VB.NET and C#) § Over 100 SQL Server + IIS Instances § Handled typical traffic of ~140 million searches per day § Average Response Time of 2.5 Seconds § Hardware upgraded as much as possible (e.g. SSDs) § Various database optimisations considered § Search-specific “cache” tables § In-memory Tables in SQL Server. § Infrastructure cost too high and reaching diminishing returns In-Memory Computing Summit, London • 25-26 June 2018

New Search-Grid Overview § Server / Cache Nodes Search Requests over HTTP Apache Ignite embedded in Spring MVC service § Cluster with Fully Replicated Caches § Most Caches Off-Heap § Jetty Process consumes around 60GB memory, including a 20GB JVM heap. § Ignite Caches Loaded from SQL Database § (with no further DB at “Search-Time”) Requests received via Embedded Jetty and processed by an Ignite § Service 20 nodes handling ~300 million searches § Ignite Update § Update Client Nodes Client Subscribes to a Message Queue § ~200k updates intraday § Updates for Availability, Rates, Static Data etc Updates Caches using a combination of Services and Ignite Data § Streamers Message Bus Updates with no visible impact on Search Process § In-Memory Computing Summit, London • 25-26 June 2018

Overall Architecture In-Memory Computing Summit, London • 25-26 June 2018

Search-Grid Internals § ~ 50 Caches § Fully Replicated § Most are Off-Heap § Cache Queries § Direct key based access where possible § SQL Fields and Indexes only when SQL Queries are necessary § Search Request § Processed by an Ignite Service § SQL Fields and Indexes only when SQL Queries are necessary § Threads managed by Ignite Services Pool § Search processed using a Single thread on a Single Node § This allows the system to be scaled up linearly In-Memory Computing Summit, London • 25-26 June 2018

Deployment § Deployment tested on § Physical Hosts § VM / Cloud Providers: AWS, Azure, Rackspace § Zero down-time Cluster deployment & restart § Starting new nodes on a separate cluster (blue/green) § Fully automated – orchestrated using Ansible § Adjusting Cluster to match Traffic Volume § Cache Nodes can be added or removed to match Traffic Volume § Caches will rebalance onto new nodes § The Event mechanism can be used to determine when all caches are rebalanced In-Memory Computing Summit, London • 25-26 June 2018

Successes & Challenges In-Memory Computing Summit, London • 25-26 June 2018

Performance § Load Test on 4 Nodes § AWS m4.4xlarge § 16 vCPU (2.3GHz XeonE5-2686) § Request Injection § 8 JMeter Injector nodes § 320 requests/sec at each step § Measurements Overview § Can sustain 960 requests / second without breaching 1-second SLA red line for 99 th % § Average response time: ~20ms 99 th Percentile: ~270ms § § Requests start queuing up beyond this rate In-Memory Computing Summit, London • 25-26 June 2018

Migration Gains § 90% reduction in infrastructure § 90% reduction in Response Time § Faster Response-Time enables new use-cases to be considered for the search process § Linearly Scalable by adding new nodes § Predictability makes infrastructure / capacity planning easier § Open Source grid-technology running on Linux § Aides quick and easy provisioning of ad-hoc Dev / Test environments § Makes it easier to have a DevOps process § New Development Processes (BDD, TDD, CI/CD) § Visible correlation between user stories and code § Test coverage provides more confidence when making complex changes In-Memory Computing Summit, London • 25-26 June 2018

Migration Pains § Need for maintaining multiple systems in the interim period § Needs to replicate the Calculation Logic, as prices must be identical to Booking System § Implicit Rounding based on Database Field precision – Multiple Temp Tables § Existing algorithms optimised for Database Queries / Stored Procedures § API Clients change their Search pattern/behavior after noticing the improved performance § Increase Search Rate § Increase in larger region/city searches § Introducing new technology required new toolsets & processes for auxiliary functions § Replacing database based monitoring & reporting tools § Many options. Needed a bit of discovery process. In-Memory Computing Summit, London • 25-26 June 2018

Supporting Services § 3 rd -party Supplier Cacheing § A more classical implementation of a Read-through cache § Reducing load on 3 rd party partners § Smarter searches to partners based on most common search types § Native Persistence § Real-Time Statistics / Analytics § Types of searches by clients § Locations being searched § Spikes in requests by Clients / Location § Integration with 3 rd party products for detailed analytics / visualisation In-Memory Computing Summit, London • 25-26 June 2018

Technical Considerations § Working with Large JVM Heaps § Garbage Collector Benchmarking / Comparison / Tuning § Development considerations to avoid long “Stop the world” pauses § Initial Rebalancing can take a long time § Need to make considerations for zero-downtime deployments § Ignite is product with a lot of active development § Great for getting lots of new useful features § Sometimes we needed help with new features, sometimes the features need some optimisations § When we found bugs, GridGain have helped by creating versions for us containing the fixes § Professional support on these issues § Developer skillset can be more business focused compared to building a platform in-house. In-Memory Computing Summit, London • 25-26 June 2018

Questions? musaul.karim@cgconsultancy.com @musaul In-Memory Computing Summit, London • 25-26 June 2018